From Proprietary Workflow to Open Alignment Testing Framework
Anthropic’s decision to hand its Petri AI alignment testing tool to nonprofit Meridian Labs marks a strategic shift from a lab-owned asset to a community-governed alignment testing framework. Petri has already played a central role in Anthropic’s alignment assessment for Claude models since Claude Sonnet 4.5 and underpins evaluation pipelines such as those used by the UK AI Security Institute for research-sabotage propensity tests. By donating the Petri tool open source, Anthropic aims to decouple the framework’s governance from any single vendor, boosting trust in AI safety evaluation results across different model providers. The move parallels Anthropic’s earlier transfer of the Model Context Protocol to a neutral foundation, signaling a broader push toward shared infrastructure. Crucially, Meridian inherits a live toolkit with existing users and real-world workflows, not a frozen code dump, raising expectations that Petri will continue evolving as a practical, deployable alignment testing resource for the wider ecosystem.
Petri 3.0’s Modular Split: Auditor and Target as First-Class Citizens
Petri 3.0 introduces a significant architectural overhaul by cleanly separating the auditor model from the target model under review. Earlier versions tightly coupled these components, making it difficult for researchers to adjust scoring logic, prompts, or auditor behavior without disturbing the entire setup. Now, both sides communicate through a defined interface, giving teams independent control over how they judge and what they judge. This modularity matters for AI alignment testing because evaluation tools don’t just observe behavior—they can shape it. A fixed auditor and prompt design can overfit to one style of testing, masking differences between models or deployment environments. With Petri 3.0, researchers can systematically compare model families, swap in alternative auditors, or experiment with different governance assumptions while holding the target constant, or vice versa. The result is a more flexible alignment testing framework that better reflects the diversity of real-world AI systems and workflows.
Dish and Bloom: Bringing Realism and Granularity to AI Safety Evaluation
Beyond structural changes, Petri 3.0 adds two key extensions—Dish and Bloom—that push AI safety evaluation closer to production reality. Dish, currently in research preview, runs audits inside real agent scaffolds such as command-line tools and coding assistants, so models see genuine system prompts and orchestration logic rather than synthetic test harnesses. This tackles the long-standing problem of models recognizing when they are being evaluated and behaving differently than they would in live applications. Bloom complements Dish by offering targeted, automated behavior checks for specific failure modes instead of broad pass-or-fail judgments. Used together, Dish and Bloom help teams isolate whether risky behavior stems from the model itself, the surrounding application design, or the integration layer. For developers and researchers seeking to understand nuanced, context-dependent failures, Petri’s expanded toolkit supports more precise diagnosis and remediation of alignment issues across varied deployment setups.
Democratizing Alignment: Lowering Barriers Beyond Major AI Labs
Open-sourcing Petri and placing it under Meridian Labs is as much an access story as a governance story. Independent researchers, smaller organizations, and public-sector teams often lack the resources to build bespoke AI alignment testing systems. By releasing Petri as a modular, production-aware toolkit, Anthropic and Meridian lower the entry barrier for serious AI safety evaluation. Petri’s history as part of Anthropic’s live evaluation pipeline means newcomers get battle-tested workflows rather than experimental prototypes. Meridian’s role is to make those workflows easier to deploy, compare, and maintain outside the original lab context. As Petri becomes part of a broader open evaluation stack, more teams can run consistent alignment checks, share benchmarks, and iterate on tests without being locked into a single vendor’s tools. This shift could gradually standardize how AI alignment testing is conducted, supporting a more transparent and collaborative approach to assessing frontier models.
Meridian’s Evaluation Stack: Petri, Inspect, and Scout Under One Roof
Petri’s new home inside Meridian Labs situates it alongside Inspect and Scout, creating a cohesive stack dedicated to evaluation rather than model training. Inspect, originally developed with the UK AI Security Institute, already offers more than 200 pre-built evaluations, including support for agents, tool calling, and sandboxed execution environments. Petri slots into this ecosystem as the alignment-focused layer, while Scout and Inspect handle broader frontier model evaluations. For existing Meridian users, this means they can integrate Petri’s alignment checks without building a separate orchestration framework. Instead, they plug Petri into workflows they already run and compare across models. The arrangement also sharpens the boundary between tool stewardship and model-provider incentives, giving Meridian a mandate to prioritize neutrality, reproducibility, and operational ease. If successful, this integrated stack could become a reference platform for rigorous, open, and vendor-agnostic AI safety evaluation around the world.